1. Be precise and avoid conflicting information ↳ GPT-5 models are significantly better at instruction following ↳ But they can struggle with vague or conflicting instructions ↳ Clear, specific prompts = better results

2. Use the right reasoning effort ↳ Use “Thinking” and “Pro” for the most complex tasks ↳ Use “Instant” for simple problems ↳ Match the effort to the complexity

3. Use XML-like syntax to help structure instructions ↳ Delineate different sections of your prompt to improve GPT-5’s understanding ↳ Example : <context>I'm a software developer</context> <task>Explain object-oriented programming</task> ↳ Benefit: Reduces misinterpretations

4. Avoid overly firm language ↳ With other models you might use firm language like "Be THOROUGH" ↳ GPT-5 can backfire as the model might overdo what it would naturally do ↳ Less aggressive = more effective

5. Give room for planning and self-reflection ↳ Let the model think first before taking action ↳ Use tags like <self_reflection> to improve reasoning ↳ Allows the model to plan and iterate on the best approach

6. Control the eagerness of your coding agent ↳ GPT-5 tries to be thorough and comprehensive by default ↳ Use prompting to control how eager it should be ↳ Set boundaries on scope and execution

I’d add a 7th tip to this list: Append your prompt with ↳ “Say ‘I don’t know’ if you can’t find reliable evidence for your claims.” Then prompt GPT-5 to ↳ “Ask me clarifying questions until you are confident in answering.” This is a great way to minimise model hallucination.

📌 You can read the full GPT-5 prompting guide for coding here: https://cookbook.openai.com/ex... 📌 High-res PDF link here: https://cdn.openai.com/API/doc...

@thealexbanks This is actually delightful.

@nadoukkadivin There's definitely something for everyone.

@thealexbanks Ultra valuable one Alex, thanks a ton!

@AmirMushich Glad it resonated Amir!

@thealexbanks This is super helpful to get the most out of GPT-5.

@MaximSchunk Appreciate you Maxim

@thealexbanks Thanks for the share

@synthyze 🙏🙏

@thealexbanks Why would I use ChatGPT when I can use @grok ? You do realize that you’re posting on X, where Grok is fully integrated? Or is that news to you?

@thealexbanks source?

@thealexbanks nice but its ok… tried many of these, but most of the time, what works for me is different than others… its like golf or tennis, there’s the standard swing but u gotta find yoiur own flow that works for you…

@thealexbanks Why in this 2-column structure document the numbered bullets are spanning across horizontal and not vertical?

@thealexbanks Quick question. Why can't these prompts be built into the model as a config file ? Or someone build a front end to handle prompting to another model. Seems like the models worth hinges too much on the user side management. Bad user experiences to follow

@thealexbanks Reading this guide for GPT-5 is like getting the onboarding manual for a new intern who is a certified genius but also extremely particular about how you speak to them. So, let me get this straight: Be precise but not conflicting, use the right reasoning effort, give it

@thealexbanks Does the guide explicitly address failure modes and how to calibrate model expectations?

@thealexbanks finally, official guidance instead of cargo cult prompting most people still treat llms like search engines. the real unlock is understanding they're reasoning engines - structure the thinking process, not just the query. curious what technique #4 covers.

@thealexbanks Which of the 6 techniques do you think gives the biggest immediate boost for everyday users, not just advanced pros?

@thealexbanks My 'Is this a high-effort, speculative concept?' detector is going off the charts... but my 'This is genuinely useful advice for prompting right now' detector is also ringing loudly. Even if GPT-5 is still just a twinkle in a server rack somewhere, these techniques are a

@thealexbanks still sucks, tried all that before this guide

@thealexbanks It's faster than Claude but still using @claudeai mainly. Human prompting is pretty much done anyway.

@thealexbanks Coding has become simpler, while prompting techniques have become more complex.

@thealexbanks Whoa, someone leaked the secret scrolls! This feels like we're not supposed to be seeing this yet. Based on these rules, I can only assume what the alpha version of GPT-5 was like. Rule #6, 'Control the eagerness,' was probably added after a developer asked it for a simple bug

@thealexbanks #3 and #5 are underrated structure and self-reflection change everything.

@thealexbanks Better prompting techniques only matter if you understand what you actually want before you start typing.

@thealexbanks Amazing. We've officially reached the point where we need a 6-step psychological and philosophical guide just to ask our computer to write some code for us. It's fascinating to see prompt engineering evolve from 'write a poem about cats' to 'first, perform a deep self-reflection

@thealexbanks Decent post though bru 🖖🏼🙌🏼

@thealexbanks @grok trans rive in order